Limits

These are the limits that can stop an action. Settings → Billing shows your current plan and usage; Settings → Storage shows saved-state usage by project and the separately reported plan allocation.

Branches

The cap is hard: a new branch is refused instead of evicting another. The count includes every non-deleted branch across the project's notebooks, regardless of status. Delete a leaf branch to free a slot.

You cannot delete the primary branch or a branch that still has child forks. Deleting the active non-primary branch from the branch selector switches away first. Branches can be cut only from the current tip. See Branches.

Sandboxes, GPUs, and RAM

A CPU sandbox is fixed at 8 vCPUs and 8 GiB. The higher RAM values are selectable system-RAM caps for GPU sandboxes.

A GPU sandbox has an absolute lifetime of 24 hours that a running cell cannot extend. Write long-training checkpoints to workspace files, then complete a boundary or export them so a later sandbox can continue safely.

Sandbox types—Base, Machine learning, Documents & OCR, and Web—are available on every plan. The type locks after the first successful cell execution. See Runtimes & sandboxes.

Saved kernel state and storage

This storage volume counts saved kernel checkpoints only, across projects you own. It does not count notebooks, uploads, datasets, workspace files, outputs, package lists, or projects someone else owns.

One branch checkpoint can be at most 4 GiB. If a namespace is larger, the cell still succeeds but that new checkpoint cannot be saved. Clusy shows a durability notice; write large frames or model weights to a workspace file, complete a boundary or export the file, and re-run the cells needed to rebuild the in-memory object.

Objects Python cannot pickle—open files, live database connections, CUDA contexts, and similar handles—are excluded at any size. Other variables still save, and the notice names excluded objects when possible.

If a new checkpoint slot is refused, Clusy shows a Storage is full dialog. An existing branch can still replace its current checkpoint so running work is not stranded. Storage has no pay-as-you-go overage.

Use Settings → Storage to see the largest projects. Clearing a project's saved kernel state is equivalent to restarting its primary kernel: variables are discarded, while notebooks, files, outputs, packages, and the environment remain. Branch checkpoints are reclaimed by deleting their branches.

Workspace file boundaries

Workspace files do not count against the saved-state plan volume, but capture and upload paths have their own operational limits:

Files over 100 MiB are restored when needed rather than copied into every fresh sandbox.

Timing

An interrupted cell is re-runnable; it is stopped the same way as the Stop button. Long jobs should checkpoint output to workspace files between cells, then complete a boundary or export critical artifacts before relying on those files after a runtime loss.

Cell source is capped at 50,000 characters, and captured output at 100,000 characters. Longer output is truncated for display rather than failing the run.

Sharing

A newly created share is a content snapshot: cells, the selected branch scope or full branch graph, and any included outputs. Runtime/kernel state is not included. A viewer who forks it reruns cells to rebuild variables.

What is not limited

• There is no fixed per-run spending limit. Allowance and an optional pay-as-you-go cap are the billing guardrails.
• Editing, branching, and reading output do not spend the model allowance. Sandbox time can still be metered while code runs.
• There is no plan cap on the number of projects or notebooks you can keep.